There has been an important progress these last years in speech recognition systems. The word recognition error rate went down with the arrival of deep learning methods. However, if one uses cloud speech API and integrate it inside a robotic architecture, one faces a non negligible number of wrong sentence recognition. Thus speech recognition can not be considered as solved (because many sentences out of their contexts are ambiguous). We believe that contextual solutions (i.e. adaptable and trainable on different HRI applications) have to be found. In this perspective, the way children learn language and how our brains process utterances may help us improve how robots process language. Getting inspiration from language acquisition theories and how the brain processes sentences we previously developed a neuro-inspired model of sentence processing. In this study, we investigate how this model can process different levels of abstractions as input: sequence of phonemes, seq. of words or grammatical constructions. We see that even if the model was only tested on grammatical constructions before, it has better performances with words and phonemes inputs.